In-Hospital Cardiac Arrest Detection Performance Analysis and Comparison on Effective Feature Selection

Abstract

Background

How to reduce the occurrence of in-hospital cardiac arrest (IHCA), screen potential IHCA patients, and advance the treatment of IHCA are urgent problems to be solved in clinic. In this study, we tried to develop a model to predict whether patients will develop IHCA based on the data of patients who have just been admitted to hospital and evaluate the influence of different feature selection methods on machine learning (ML) models.

Methods and Results

A total of 25 149 patients were included in the study; 320 developed IHCA. We chose three feature selection methods (Student's t-test and Chi-square test, regression analysis and correlation analysis) and four ML models (AdaBoost, XGBoost, Random Forest, and Logistic Regression). Each ML model was trained and evaluated using raw and feature-selected data; as a result, we got 16 models. AUROC, AUPRC, accuracy, recall, precision, and specificity are used to evaluate the model. The XGBoost model has the best performance with an AUROC of 0.987 (95% CI 0.984–0.988), an AUPRC of 0.763, an accuracy of 0.992, a recall of 0.695, a precision of 0.723, and a specificity of 0.996. The most significant predictors are age, albumin, sinus arrhythmia, activated partial thromboplastin time, and protein.

Conclusions

Different feature selection methods have different effects on different ML models. The predictive model developed using the XGBoost algorithm is the best predictor of whether patients will develop IHCA.